Zapier Audit Saves $4,788/Year, AI Tooling Reshapes SaaS and Startup Playbooks

Key Takeaways

1. Noah Kagan cut his Zapier bill from $4,788/year to $0 after auditing 123 automations and finding 63 unused and 2 doing 88% of the work.
2. Tony Dinh replaced a translation SaaS with a built-in GPT-Luna 6 tool, saving $124/month across 2,705 keys and 16 languages.
3. Alex Lieberman published an AI portfolio matrix and readiness checklist for deciding which AI projects to fund and when to invest in enablers.
4. A new analysis of mechanistic interpretability argues neural networks are largely readable, with 90% of audited Qwen3-32B neurons having one primary function.
5. Roman's Career Hound hit $28k/month with a hard paywall and a repeatable content format, showing distribution beats product novelty.
6. Acquire.com listed four SaaS businesses with TTM revenue from $184k to $6.7M, plus a free Listing Scorecard for sellers.

1. AI Tooling and Cost Discipline

1. Noah Kagan stopped using Zapier after 10 years: of 123 automations, 63 had not run in a month and 2 handled 88% of the workload, mostly form-to-Slack and HubSpot-to-Sheets connections. He rebuilt one with Claude, moved the rest into already-paid tools, and cut annual spend from $4,788 to $0. His audit method: export Zapier task history, have Claude or ChatGPT list each automation, owner, and 30-day run count, flag unused ones, identify the few carrying the load, and match each to a paid alternative with a dated cancellation plan. — via 1
2. Tony Dinh said his motivation was reducing dependency rather than cutting costs, with savings as a side benefit. His app replaced an i18n translation SaaS with a built-in tool covering 2,705 translation keys and 16 languages; GPT-Luna 6 costs about $0.75 and he said translations are more accurate and context-aware, saving $124 per month. — via 1
3. Alex Lieberman proposed an "AI portfolio matrix" that sorts AI opportunities by return size and current executability into four buckets: immediate acceleration, investment to unlock, quick wins, and deprioritize. His four-step process runs from goals to inventory to rules to sequencing. He stressed that every pilot needs a business owner, verifiable results, and a cost estimate; small fixes can proceed directly if cheap, owned, and effective within weeks; go/no-go criteria must be written before launch; and early evidence should show faster, cheaper, or better without quality loss, daily employee use, gains actually caused by AI, and valid cost-benefit assumptions. Readiness checks cover clear value, understandable and measurable process, accessible data, deployment and support, target-user involvement, team capacity to change workflows, owner accountability, and legal guardrails. The first wave should come from the immediate-acceleration quadrant, ranked by success probability then return, with quick wins in parallel; when a gap blocks multiple high-return projects, invest in enabling work even if it produces no direct return. — via 1

2. AI Research and Interpretability

1. Alex Lieberman relayed and interpreted Nick Cammarata's new article on mechanistic interpretability, which argues neural networks are broadly readable and that the "tangled mess" narrative is exaggerated because of poor old tools and model design. The four-step argument: networks use only one learning principle, modern tools exceed expectations, understanding compounds, and direct research may already be close to what is needed. Evidence cited includes cat/car neurons actually detecting sharp V shapes, a car embedded in a dog neuron actually being a pose neuron, only 0-5% of InceptionV1's final layer being interpretable in 2014 due to model and tool issues, reducing neurons from 480 to 10 producing broader categories rather than superposition, sparse autoencoders finding content that is about 88% duplicates of existing neurons, and an audit of 1,000 Qwen3-32B neurons showing 90% with one primary function, 7% with two related functions, and 2-4% randomly mixed. The article's acknowledged weak points: language-model evidence is thin, the detailed technical paper is not yet published, "90% one function" depends on how a function is defined, and weak signals still lack a reliable test, which is where superposition supporters say hidden concepts live. — via 1

3. Startup Growth and Marketing Playbooks

1. Roman built Career Hound over two years by scraping job listings directly from company websites, with no free tier and a hard paywall at $7/week, $20/month, or $50/year. It reached $28k in revenue last month, 8,000+ users, 779 active subscriptions, and 11.7 million social views in the past 30 days. Starter Story argued the product itself is not special; the difference is Roman's marketing ability, shaped by studying psychology and five summers selling alarm systems door-to-door, which taught him to manage creators and use content to trigger emotion. His methods include hours of "content gold mining" on Instagram, Reddit, and TikTok; ignoring weekly trends for evergreen formats; and a "Companies hiring remote worldwide" list format that has run for two years. He prefers hiring creators over 50, believing gray hair adds subconscious credibility to job-search advice, pays about $20 per video, DMs people already making content for fun, and gives minimal instructions so they do no research. His stack is lean: Shortimize, PostHog, WhatsApp, Clarity, Hetzner, Postmark, Google Suite, and Cloudflare. The takeaway: you do not need a new idea, you need a mature category plus continuous marketing experiments; Roman said "creating a profitable business is actually inevitable." — via 1
2. Acquire.com listed four new businesses: a mobile AI learning app that turns course materials into personalized exam prep with $38.7k TTM revenue asking $100k; a digital fitness brand with about $12.2M cumulative Stripe receipts, ~$6.9k subscription MRR, $184k TTM revenue asking $200k; an AI tool for VA claims positioned as "TurboTax for VA claims" that finds benefits and files in minutes, with $6.7M TTM revenue asking $10M; and a web-to-print SaaS for printers with $258.7k TTM revenue asking $699.4k. The platform also launched a free Listing Scorecard that grades listings on buyer priorities and shows their position. — via 1 2 3 4 5
3. Acquire.com shared Paresh Deshmukh and Chirag's path: they built their first SaaS full-time, took over two months to get the first 10 customers, reached 100k+ installs and $12k MRR, and sold on Acquire.com at that MRR. They then founded BoloForms, which reached nearly $1M ARR and over 5 million users, and Paresh is now back on the platform looking to acquire SaaS companies. — via 1

4. Founder Perspectives and Market Signals

1. Codie Sanchez said the top predictor of success she has seen is firsthand experience, and the most successful founders in her VC portfolio had worked on the problem they solved for over 10 years. She quoted a founder of an $11B company calling managers a waste of space and linked to an alternative hiring approach. She argued that if a startup founder is not all-in, even a little delusional, the company cannot win, and that this delusion can produce bestsellers and big businesses. — via 1 2 3
2. Nick Huber posted a series of controversial personal views: married adult men should not climb ladders to install Christmas lights; bigger cars are safer and families should use SUVs, with crash ratings called a scam; pit bulls are a public nuisance; politics wastes time and is negatively correlated with happiness. He said making money is hard, $100k a year is a lot, Twitter makes it look easy, sports betting is a social disaster quietly ruining many men, and waiting until 35 to consider children is a huge mistake. He called SEC and Big Ten schools the new Ivy League, said religion is returning and those who take it seriously live better, and said marrying a doctor or lawyer brings high status but low enjoyment and little flexibility. He argued vaping and marijuana are better than alcohol for the body but worse for career and life, that 75 vaccines will be proven a net negative, and that technology will not advance much further, with the physical world basically unchanged in 20 years. He predicted autonomous driving is 20 years away and Waymo will not leave 25 mph cities, VR will fail for lack of demand, 99% of people have no use for AI, AI has already delivered 90% of its potential, and it is a huge waste of electricity. He called Web3 a joke, crypto worthless, Bitcoin pure speculation and a waste of power solving a nonexistent problem; EVs will not go mainstream and are worse for the environment; Starlink is a miracle worth 10 times the social value of the previous five combined, and Musk should get a Nobel Prize. He said the Fed did well on rates over the past three years, nuclear is the only way forward, solar and wind are unsustainable and wasteful, oil and gas are the only viable option today and the US should invest heavily in production, and Silicon Valley influencers and tech giants do not understand what Americans need, suggesting a Walmart visit would help because most people just want to put food on the table and do not need another app for Meta glasses. — via 1
3. Ben Tossell described his AI workflow on a podcast: he calls himself non-technical and fears the terminal, yet has shipped code to production at FactoryAI. He stores agent context in folders, does not use Claude Code, and said AI makes him more willing to spend time at the computer. The episode covered learning to work with AI through "boring small projects," building a mascot animator for his personal site, making a visual dictionary to understand coding-agent terms, splitting tasks among Pi, Factory, and Codex, agreeing that faster models feel smarter and preferring to steer agents mid-work, using a shared vocabulary so humans and agents do not talk past each other, and questioning whether calling AI adoption a "skills problem" lets AI labs off the hook. He also makes personal context portable across agents and models, reduces what agents remember about him, and advocates constantly asking AI questions. He raised an open question: with personal agents, will usage converge on a single thread or multiple agents? He also relayed that native HTTPS support has been added to wasm builds, so agents can publish and host ports. — via 1 2 3
4. Levelsio relayed a newsletter tool described as "vibe-coded": it claims to be 83% cheaper than beehiiv, open-source and free to self-host, with an editor, scheduling, stats, per-issue websites, deliverability protection, and sending from your own domain. It supports two-minute imports from beehiiv, Substack, and Mailchimp, charges by emails sent rather than dormant subscribers, starts at $19/month, and has no "Powered By" badge at the bottom of emails. — via 1
5. Shaan Puri recommended a podcast for a one-hour drive this week, covering the A-B-Z framework for stopping overthinking, how your hometown may limit your growth, and the story of "Iron John." — via 1
6. Justin Welsh argued that when you are unsure of your direction, act first: write, record, publish, call, email, or network, because motivation comes from action, not thought. He said it is nearly impossible to achieve extreme success without a period of looking foolish, and relayed that the things that truly change life, such as reading, exercising, walking, and more honest conversations, are extremely boring. He also relayed that the most stressful lifestyle is a job that desperately needs you. — via 1 2 3 4